We need to talk about how LLM-generated clinical notes might make it even harder to find real information in the EMR.
Imagine you are an ICU physician accepting a patient in transfer from an outside hospital (in other words, you’re me). Your unit clerk has printed out 100 pages of clinical records from the transferring facility, which you’re rifling through, trying to piece together what actually happened to the patient. You flip past the medication administration tables, hourly vital sign readings, records of routine nursing tasks, laboratory results, laboratory results re-copied into resident notes, radiology reports, duplicated radiology reports… Ah, here! At last! The discharge summary generated by an actual, thinking human.
Now imagine your disappointment when you start reading and, a few paragraphs into the Hospital Course and Significant Findings, you realize it’s the same EMR data crafted into smart-sounding prose. It’s eerily human-like, but lacking in narrative direction. You’re not sure whether it’s making actual sense or not.
Sure, it might be an intern note (that’s a joke — I appreciate you interns). Or it might be GPT-4.
Ambient generation of clinical notes risks replacing the EMR’s last bastion of cogent thought with verbose LLM-generated drivel. We have seen EMR note automation go wrong before when we started pulling in reams of lab data and medication lists, copying forward stale content, and filling out forms with drop-down selections that at best represent a cookie-cutter, machine-readable shadow of what’s happening with the patient. Supposedly LLMs will be smarter than prior generations of automation. But that depends on us leverage them intelligently, and not repeating the mistakes of the past.
I might be saying this because I’m a curmudgeonly general surgeon, with the trained conservatism of that specialty — hey, you want your surgeon to be careful! — but I’m not against AI adoption. I’m a technology optimist, and this technology has arrived. LLM-generated notes are here to stay because physicians need the relief they offer. So my interest is in steering this technology in the right direction.
Doctors need to keep an eye on what is happening with clinical notes. There are billing-driven requirements which the ambient tools can help us solve, but let’s not forget clinical communication. From my perspective, that’s what our notes are for. Contrary to the current frenzy of AI exuberance, I assert that doctors will still want to say the important things in our own words. Soon, notes will have two parts: the automated content, and the “here’s what I really think” content.
There is pre-AI precedent for this. In hospitals and clinics where residents help us see patients, attending physicians write an attestation of the resident notes (which themselves vary in their usefulness). Often the attending writes just a sentence or two explaining what’s actually happening. At my teaching hospital, Epic now puts the attending attestation at the top of the resident note — because sometimes that’s all you have to read.
We are at the point in the generative AI hype cycle where the limitations are coming into focus. We can do a lot with the technology. But as a documentation tool, it’s not replacing doctors, not yet. It’s just helping us give CMS, UnitedHealthcare, and our hospital administrators what they want… while we focus on what matters.
Let’s make sure doctors stay in the driver’s seat. This technology is not yet our copilot, but it might be our car — meaning we’ll probably go a lot faster with it than without it.
